SHAP-based Interpretable Models for Credit Default Assessment Using Machine Learning

In recent years, the issue of credit fraud risk has garnered increased attention from the banking and financial sectors. However, prevailing credit assessment models predominantly focus on predictive outcomes, often overlooking the imperative of model interpretability. Understanding the contributions of model features and their interactions is paramount for elucidating model behavior and furnishing vital insights for model enhancement and optimization. To address this gap, this paper proposes a machine learning model leveraging SHAP for explicating credit assessment. Utilizing publicly available datasets from Lending Club, this study validates the proposed model against four industry-standard machine learning approaches: SVM, MLP, XGBoost, and LightGBM. Experimental findings unveil feature importance rankings and elucidate relationships between features and target variables. Notably, the study identifies the predominant roles of loan interest rates and credit policies in credit fraud assessment within the dataset and endeavors to uncover interactions within individual key features. The SHAP framework, as demonstrated, holds promise for informing the design and construction of future credit risk assessment models, thereby bolstering support for financial decision-making and risk management endeavors.